Built by Xiaomi's MiMo team under Fuli Luo, a former DeepSeek researcher, MiMo-V2.6-Pro is MIT-licensed and self-hostable in BF16, FP8 or INT8. It suits teams wanting an open, agent-ready alternative to closed frontier labs, though Xiaomi has not published a system card or training-data cutoff.
MiMo-V2.6-Pro is Xiaomi's flagship open-weights model, released September 22, 2026, with 1.02 trillion total parameters and 42 billion active per token. It scored 46 on the Artificial Analysis Intelligence Index, the top mark among open-weights models at release, with a 1,048,576-token context window.
Where it sits
- $0.544/M$ per 1M tokensBlended price (3:1)Lower is better#18 / 64peer median $1.70/Mvendor price, checked by HokAI
- 130 tok/stokens/sOutput speedHigher is better#13 / 39peer median 90 tok/scited: Artificial Analysis
- --% solvedSWE-bench VerifiedHigher is better-- / 28peer median 78.3%per source, see benchmark scores
- --% correctGPQA DiamondHigher is better-- / 44peer median 88.3%per source, see benchmark scores
Cheaper than 71% of the 64 GA models with a published price, and rank 13 of 39 on output speed as cited from Artificial Analysis.
Ranks are against GA models on HokAI that publish the same figure; ties share a rank.
Provider: Xiaomi · Family: MiMo V2.6
Context window: 1,048,576 tokens · Max output: 131,072
Input modalities: text, image, video, audio · Output: text
About MiMo-V2.6-Pro
MiMo-V2.6-Pro is the flagship model in Xiaomi's MiMo family, built under lead researcher Fuli Luo, a former DeepSeek and Alibaba DAMO Academy researcher, and released on September 22, 2026. It is a sparse Mixture-of-Experts transformer with 1.02 trillion total parameters and 42 billion active per token, spread across 70 transformer layers (60 sliding-window, 10 global-attention) with a hidden size of 6144 and 384 routed experts, 8 active per token. It ships alongside a smaller sibling, MiMo-V2.6-Flash (309B total / 15B active), and a latency-optimized MiMo-V2.6-Pro-UltraSpeed variant from the same checkpoint, succeeding MiMo-V2.5-Pro in Xiaomi's push, backed by a committed $8.7 billion in AI investment over three years, to make MiMo a credible open-weights alternative to closed frontier labs.
On the Artificial Analysis Intelligence Index (v4.3, a ten-evaluation composite covering reasoning, knowledge, math and coding), MiMo-V2.6-Pro scores 46, reported by Artificial Analysis as the top score among open-weights models at release, at roughly $0.13 per Intelligence Index task. On Xiaomi's own reported evaluations, the model scores 82.0 on OSWorld-Verified (computer-use), 71.9 on DeepSWE v1.1 (agentic coding), 94.0 on CyberGym (cybersecurity) and 72.3 on MiMo VisualCoding. Xiaomi's release materials describe these scores as on par with proprietary leaders including GPT-5.6 Sol and Claude Opus 5 on most agent benchmarks, a comparison HokAI has not independently verified.
The model carries a 1,048,576-token context window with a 131,072-token max completion, sized for long repositories and multi-session agent runs, and uses a 5-layer speculative decoder for multi-token prediction. Artificial Analysis independently measured 129.7 output tokens per second and a 2.17-second time to first token on the standard endpoint.
MiMo-V2.6-Pro is omni-modal on input (text, image, video, audio), backed by a 681-million-parameter vision encoder and roughly 435 million parameters of audio encoders. Output is text only; there is no native audio or image generation. It supports tool calls, structured output and an opt-in reasoning trace via OpenRouter's "reasoning" and "include_reasoning" parameters, making it usable as a function-calling agent backend.
Xiaomi prices MiMo-V2.6-Pro at $0.435 per million input tokens and $0.87 per million output tokens through its own mimo.mi.com console and resellers like OpenRouter, with cached input at roughly $0.0036 per million tokens, a 99% discount over fresh input. The UltraSpeed variant, from the same checkpoint, is priced roughly tenfold higher in exchange for substantially faster generation; see the pricing table below for exact figures.
Weights are published on Hugging Face (XiaomiMiMo/MiMo-V2.6-Pro-RL) and GitHub as safetensors in BF16, FP32, FP8 and INT8, letting self-hosters pick a precision for their hardware; no VRAM chart is published. Hosted access runs through Xiaomi's own console and OpenRouter; no AWS Bedrock, Google Vertex or Azure listing has surfaced.
Xiaomi describes an "Aligned RL" approach pairing large-scale reinforcement learning with environment hardening, adversarial screening and verifier cross-checks against reward hacking, plus a self-correction step where the model rewrites its own misaligned reasoning. No system card, named red-teaming partners, or refusal-rate benchmark has been published for MiMo, a thinner disclosure than OpenAI, Anthropic or Google DeepMind give their flagships.
MiMo-V2.6-Pro suits teams wanting an open-weights, self-hostable frontier model for agentic coding, computer-use and long-context work, and anyone price-sensitive, since pricing undercuts most proprietary flagships by roughly an order of magnitude. It is a weaker fit where audited safety documentation, compliance certifications or guaranteed API stability matter; DeepSeek-V4 and Alibaba's Qwen3 are the closest open-weights alternatives to compare against.
Xiaomi has not disclosed a training-data cutoff or the training corpus behind MiMo-V2.6-Pro. The model carries the MIT license, permitting commercial use, modification and redistribution without royalty. Retention depends on deployment: self-hosted inference keeps only what the operator logs, while the hosted API's retention policy is not published in English documentation found during this research.
MiMo-V2.6-Pro follows the March 2026 launch of MiMo-V2-Pro, MiMo-V2-Omni and MiMo-V2-TTS, the mid-2026 MiMo-V2.5-Pro generation, and June 2026's open-sourcing of MiMo Code. In the week before release, the MiMo team livestreamed the RL training run behind V2.6 in public at mimo.xiaomi.com/rl, publishing reward curves and a cost counter that passed a combined $3 million in spend across the Pro and Flash checkpoints by September 19, 2026, two days before release.
Pricing
Pricing confirmed via OpenRouter's public model listing and Xiaomi's mimo.mi.com pay-as-you-go console. The MiMo-V2.6-Pro-UltraSpeed variant, built from the same checkpoint, costs 10x more per token ($4.35 input / $8.70 output) for faster output.
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.013 | $0.0009 | $0.014 |
| Support reply | $0.0009 | $0.0003 | $0.0011 |
| One coding agent run | $0.087 | $0.017 | $0.104 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Key Features
- 1M-Token Context Window: 1,048,576-token context with a 131,072-token max output, sized for long repositories, extended tool traces and multi-session agent runs.
- Omni-Modal Input: Accepts text, image, video and audio via dedicated 681M-parameter vision and roughly 435M-parameter audio encoders; output remains text-only.
- MIT-Licensed Open Weights: Full weights published on Hugging Face and GitHub under the MIT license in BF16, FP8 and INT8 formats for self-hosting at multiple precision points.
- UltraSpeed Serving Option: A latency-optimized endpoint built from the same 1T-parameter checkpoint, priced well above standard for substantially faster generation.
- Toggleable Reasoning Traces: Supports an opt-in reasoning / include_reasoning API parameter for visible chain-of-thought on complex agentic and coding tasks.
Pros
- Tops the Artificial Analysis Intelligence Index among open-weights models at a score of 46, at roughly $0.13 per Intelligence Index task.
- MIT-licensed weights across multiple precision formats let teams self-host at the size that fits their hardware.
- Input token pricing undercuts most proprietary flagship models by roughly an order of magnitude.
Cons
- No published system card, named red-team partners, or refusal-rate benchmark.
- Output is text only, with no native audio or image generation despite omni-modal input support.
- No disclosed training-data cutoff date, making recency of world knowledge hard to judge.
Benchmarks
- OSWorld Verified: 82% vendor-reported · 22 Sep 2026 — Tasks completed by operating a real desktop, % solved.
- AA Intelligence Index: 46 cited: Artificial Analysis · 22 Sep 2026 — Composite of 10 evaluations run by Artificial Analysis, 0 to 100.
- AA blended price: $0.18/M cited: Artificial Analysis · 22 Sep 2026 — Price per 1M tokens at a 3:1 input to output blend, as listed by Artificial Analysis.
- Output speed: 130 tok/s cited: Artificial Analysis · 22 Sep 2026 — Median tokens written per second as measured by Artificial Analysis.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Frequently Asked Questions
How much does MiMo-V2.6-Pro cost per 1M tokens?
MiMo-V2.6-Pro costs $0.435 per million input tokens and $0.87 per million output tokens through Xiaomi's own mimo.mi.com console or OpenRouter, with cached input around $0.0036 per million tokens, a 99% discount. That undercuts proprietary flagships like GPT-5.6 Sol or Claude Opus 5 by roughly an order of magnitude per token. The UltraSpeed variant, built from the same checkpoint, costs 10x more per token for faster serving.
How does MiMo-V2.6-Pro compare on benchmarks vs DeepSeek V4?
MiMo-V2.6-Pro scored 46 on the Artificial Analysis Intelligence Index, reported as the top mark among open-weights models at its September 2026 release, ahead of DeepSeek's V4 line on that composite. Xiaomi's own materials also report 82.0 on OSWorld-Verified for computer-use tasks and 94.0 on CyberGym for cybersecurity, though these are Xiaomi-reported figures rather than independently re-run scores.
Is MiMo-V2.6-Pro open source or proprietary?
MiMo-V2.6-Pro is released under the MIT license with full weights published on Hugging Face and GitHub in BF16, FP32, FP8 and INT8 formats, allowing commercial use, modification and self-hosting without royalty. It can also be reached as a hosted API through Xiaomi's own console or resellers like OpenRouter.
Does MiMo-V2.6-Pro train on user data?
Xiaomi has not published a data-retention or training-on-inputs policy for the hosted MiMo API in English-language documentation found during this research. Self-hosted deployments retain only what the operator's own infrastructure logs, since no data leaves that environment in that mode.
Who is MiMo-V2.6-Pro best for and who should avoid it?
It suits teams wanting a self-hostable, MIT-licensed model for agentic coding or long-context (1M-token) workloads at a fraction of proprietary pricing. Teams that need audited safety documentation, named red-team partners or enterprise compliance certifications should look elsewhere, since Xiaomi has not published any of the three for MiMo as of this release.